Network Status Predictor to support dynamic scheduling in network-based computing systems

The management of networks has often been ignored in network-based computing systems due to the difficulty of estimating application programs' network latency and bandwidth requirements, and the difficulty of predicting the system network load. To help address this deficiency, and thereby support dynamic network resource scheduling, we propose the Network Status Predictor (NSP). This tool is a general and extensible network load monitor that introduces lower and upper latency prediction bounds. We evaluate its ability to dynamically predict TCP/IP end-to-end latency with varying network loads using a cluster of SGI multiprocessors interconnected with a Fibre Channel network. Our results show that a combination of numerical predictors can be dynamically selected based on the network's recent state to produce better predictions than when using a single predictor alone.